[Bug c/34174] New: gcc produces erroneous asm for movdi

[Bug c/34174] New: gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

Compiling the testcase for BUG #27386 with the fr30-elf crosscompiler
shows that while reloading memory operands to registers the source adress of
the memory operand is changed.

testcase:

typedef unsigned long long uint64_t;

uint64_t a, b, c;

int
foo(uint64_t x, uint64_t y, uint64_t z, int i)
{
        a = x;
        b = y;
        c = z;
        return 2 * i;
}

int main( void )
{
        return foo(1234512345123ull,
                   3456734567345ull,
                   7897897897897ull,
                   42);


}

Looking at the assembler output which should store argument x in a

        ldi:32  a, r3   
        ldi:8   #248, r1
        extsb   r1      
        addn    fp, r1  
        ld      @r1, r1 
        mov     r1, r2    <-- *
        addn    4, r2   
        ld      @r2, r2 
        st      r1, @r3 
        st      r1, @-r15       
        mov     r3, r1  
        addn    4, r1   
        st      r2, @r1 

At the position marked with * the compiler probably wants the adress of a which
was stored in r1 but r1 was changed by the previous ld instruction.


-- 
           Summary: gcc produces erroneous asm for movdi
           Product: gcc
           Version: 4.2.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: markus dot heigl at fme dot fujitsu dot com
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: fr30-unknown-elf


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug c/34174] gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

------- Comment #2 from markus dot heigl at fme dot fujitsu dot com  2007-11-21 12:37 -------
Created an attachment (id=14591)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14591&action=view)
The output asm file with sibling


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug c/34174] gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

------- Comment #1 from markus dot heigl at fme dot fujitsu dot com  2007-11-21 12:36 -------
Created an attachment (id=14590)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14590&action=view)
The used testcase


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #3 from rask at gcc dot gnu dot org  2007-11-21 15:40 -------
Created an attachment (id=14592)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14592&action=view)
patch v1 for GCC 4.3

Please use -dp when posting asm output because it makes it easier to see what
is going on. Please also state the compiler options needed to reproduce the
bug.
Anyway, confirmed on trunk with revision 130319 with -O0. It is the usual
problem of a target which defines movdi patterns when it shouldn't and the fix
is to just delete the crap, which the attached patch does. It even saves an
instruction.

Before (function foo):
...
        ldi:32  a, r3   ; 13    movsi_internal/4        [length = 6]
        ldi:8   #248, r1        ; 52    movsi_internal/1        [length = 2]
        extsb   r1      ; 53    extendqisi2             [length = 2]
        addn    fp, r1  ; 16    addsi_regs              [length = 2]
        ld      @r1, r1 ; 74    movsi_internal/7        [length = 2]
        mov     r1, r2  ; 75    movsi_internal/5        [length = 2]
        addn    4, r2   ; 76    addsi_small_int/1       [length = 2]
        ld      @r2, r2 ; 77    movsi_internal/7        [length = 2]
        st      r1, @r3 ; 78    movsi_internal/6        [length = 2]
        mov     r3, r0  ; 79    movsi_internal/5        [length = 2]
        addn    4, r0   ; 80    addsi_small_int/1       [length = 2]
        st      r2, @r0 ; 81    movsi_internal/6        [length = 2]
...
After:
        ldi:32  a, r3   ; 19    movsi_internal/4        [length = 6]
        ldi:8   #248, r1        ; 77    movsi_internal/1        [length = 2]
        extsb   r1      ; 78    extendqisi2             [length = 2]
        addn    fp, r1  ; 22    addsi_regs              [length = 2]
        ld      @r1, r2 ; 23    movsi_internal/7        [length = 2]
        st      r2, @r3 ; 24    movsi_internal/6        [length = 2]
        mov     r3, r2  ; 82    movsi_internal/5        [length = 2]
        addn    4, r2   ; 26    addsi_small_int/1       [length = 2]
        addn    4, r1   ; 28    addsi_small_int/1       [length = 2]
        ld      @r1, r1 ; 29    movsi_internal/7        [length = 2]
        st      r1, @r2 ; 30    movsi_internal/6        [length = 2]

Please try this patch with GCC 4.2.2. Also, do you happen to know of a
simulator for fr30?


-- 

rask at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |rask at gcc dot gnu dot org
                   |dot org                     |
             Status|UNCONFIRMED                 |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

------- Comment #4 from markus dot heigl at fme dot fujitsu dot com  2007-11-21 21:46 -------
I will test it tomorrow when I'm at work again.

There are simulators available from Greenhills and an Instruction set simulator
from Cygnus. But that are both commercial products.

Our own IDE Softune Workbench has a simulator builtin but it is written for
Windows.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 544 bytes --]



------- Comment #5 from markus dot heigl at fme dot fujitsu dot com  2007-11-22 09:39 -------
When trying to build a cross compiler with this patch I get following error:

dp-bit.c: In function â__muldf3â:
dp-bit.c:964: internal compiler error: in change_address_1, at emit-rtl.c:1783

build options:

../gcc-4.2.2/configure --target=fr30-elf
--prefix=/fme/users/mhaigl/unstable-gcc --with-gnu-as --with-gnu-ld
--without-headers --with-newlib --enable-languages=c --disable-shared


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #6 from rask at gcc dot gnu dot org  2007-11-22 12:31 -------
Created an attachment (id=14605)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14605&action=view)
patch v1 for GCC 4.2.2

Here's a different patch which hopefully doesn't ICE on GCC 4.2.2.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 655 bytes --]



------- Comment #7 from markus dot heigl at fme dot fujitsu dot com  2007-11-22 12:50 -------
Building with this patch produces following error.
Same build options as last time.

../../gcc-4.2.2/gcc/libgcc2.c: In function â__udivmoddi4â:
../../gcc-4.2.2/gcc/libgcc2.c:1054: error: insn does not satisfy its
constraints:
(insn 908 907 903 52 ../../gcc-4.2.2/gcc/libgcc2.c:962 (set (reg:DI 2 r2)
        (mem/c:DI (reg:SI 2 r2) [27 rr+0 S8 A32])) 7 {*movdi_insn} (nil)
    (nil))
../../gcc-4.2.2/gcc/libgcc2.c:1054: internal compiler error: in
reload_cse_simplify_operands, at postreload.c:392


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #8 from rask at gcc dot gnu dot org  2007-11-22 13:54 -------
Created an attachment (id=14607)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=14607&action=view)
patch v2 for GC 4.2.2

Weird. I don't understand why GCC 4.2.2 is having problems with that. This
patch tries to fix fr30_move_double(), which has been broken ever since it was
added seven and a half years ago. The instruction sequence now looks like this:

        ldi:32  a, r3   ; 13    movsi_internal/4        [length = 6]
        ldi:8   #248, r1        ; 52    movsi_internal/1        [length = 2]
        extsb   r1      ; 53    extendqisi2             [length = 2]
        addn    fp, r1  ; 16    addsi_regs              [length = 2]
        mov     r1, r2  ; 74    movsi_internal/5        [length = 2]
        ld      @r1, r1 ; 75    movsi_internal/7        [length = 2]
        addn    4, r2   ; 76    addsi_small_int/1       [length = 2]
        ld      @r2, r2 ; 77    movsi_internal/7        [length = 2]
        st      r1, @r3 ; 78    movsi_internal/6        [length = 2]
        mov     r3, r0  ; 79    movsi_internal/5        [length = 2]
        addn    4, r0   ; 80    addsi_small_int/1       [length = 2]
        st      r2, @r0 ; 81    movsi_internal/6        [length = 2]


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[markus dot heigl at fme dot fujitsu dot com]

------- Comment #9 from markus dot heigl at fme dot fujitsu dot com  2007-11-22 15:59 -------
Seems to work for this testcase.

I think the else path of if (reverse) should also be changed in the same way.
Do you agree?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #10 from rask at gcc dot gnu dot org  2007-11-23 01:33 -------
I think both branches of "if (reverse)" could use the exact same code, i.e.
this whole reverse/!reverse idea is bogus on fr30. Suppose our output registers
are r1 and r2 and we receive the address in rN. Then, for any N, this
instruction sequence should work:
...
mov  rN,  r2
ld   @rN, r1
addn 4,   r2
ld   @r2, r2
...
I found an fr30 simulator in GDB 5.2, so I'm testing the patch for GCC 4.3.
Because this bug is not a regression, it might not be fixed in GCC 4.2.x.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #11 from rask at gcc dot gnu dot org  2007-11-23 13:46 -------
>  dp-bit.c:964: internal compiler error: in change_address_1, at emit-rtl.c:1783

That is because it needs an offsetted address during and after reload, and the
fr30 doesn't have that. It can happen on trunk too.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[patchapp at dberlin dot org]

------- Comment #12 from patchapp at dberlin dot org  2007-11-23 17:10 -------
Subject: Bug number PR 34174

A patch for this bug has been added to the patch tracker.
The mailing list url for the patch is
http://gcc.gnu.org/ml/gcc-patches/2007-11/msg01232.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #13 from rask at gcc dot gnu dot org  2007-11-26 13:20 -------
Subject: Bug 34174

Author: rask
Date: Mon Nov 26 13:20:19 2007
New Revision: 130438

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130438
Log:
        PR target/34174
        * config/fr30/fr30.c (fr30_move_double): Sanitize mem->reg case. Copy
        the address before it is clobbered.

testsuite/
        * gcc.dg/torture/pr34174-1.c: New.

Added:
    trunk/gcc/testsuite/gcc.dg/torture/pr34174-1.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/fr30/fr30.c
    trunk/gcc/testsuite/ChangeLog


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #14 from rask at gcc dot gnu dot org  2007-11-28 01:44 -------
Subject: Bug 34174

Author: rask
Date: Wed Nov 28 01:44:10 2007
New Revision: 130489

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=130489
Log:
        Backport from mainline:
        2007-11-26  Rask Ingemann Lambertsen  <rask@sygehus.dk>

        PR target/34174
        * config/fr30/fr30.c (fr30_move_double): Sanitize mem->reg case. Copy
        the address before it is clobbered.

testsuite/
        * gcc.dg/pr34174-1.c: New.

Added:
    branches/gcc-4_2-branch/gcc/testsuite/gcc.dg/pr34174-1.c
      - copied, changed from r130438,
trunk/gcc/testsuite/gcc.dg/torture/pr34174-1.c
Modified:
    branches/gcc-4_2-branch/gcc/ChangeLog
    branches/gcc-4_2-branch/gcc/config/fr30/fr30.c
    branches/gcc-4_2-branch/gcc/testsuite/ChangeLog


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174

[Bug target/34174] gcc produces erroneous asm for movdi

[rask at gcc dot gnu dot org]

------- Comment #15 from rask at gcc dot gnu dot org  2007-11-28 01:55 -------
Fixed for both 4.2.3 and 4.3.0.


-- 

rask at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
      Known to fail|4.2.2 4.3.0                 |4.2.2
      Known to work|                            |4.2.3 4.3.0
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34174